Agentic engineering
Vision first
I set the destination, give AI a strong scaffold and let the system iterate towards the goal.
“Don't chase the models. Build the scaffold they rise into.”
Where this is heading: the five levels
Dan Shapiro maps AI-assisted development onto a ladder, from spicy autocomplete to the software factory: at the bottom you write the code; at the top you set the destination.
| Level | Stage | What it means |
|---|---|---|
| 0 | Spicy autocomplete | AI is a reference tool; the developer writes the code. |
| 1 | Coding intern | AI handles snippets and boilerplate under immediate review. |
| 2 | Junior developer | Human and AI work interactively as pair programmers. |
| 3 | Developer | AI writes most of the code, but the human reviews every change. |
| 4 | Senior developer | AI works unattended, relying on automated checks and delayed human review. |
| 5 | Software factory | People set goals; the system plans, builds, verifies and ships. |
Most teams sit around level 2 or 3, a human as the bottleneck on every change. I work at level 4 now, building towards 5.
The goal is a self-correcting loop, not a human reviewing every diff
My model comes from situational awareness: see where you are, adapt, act, check the result and correct course.
- See
- Adapt
- Act
- Check
- Correct
Stronger models make the scaffold more valuable
My answer is the rising tide. The markdown skills a team builds sharpen with every pass. Each pass improves the question; a better question and a better model compound, so stronger models make the same system more valuable.
